Intelligent IR on the World Wide Web CSC 575 Intelligent Information Retrieval

2 Intelligent IR on the World Wide Web  Web IR versus Classic IR  Web Spiders and Crawlers  Citation/hyperlink Indexing and Analysis  Intelligent Agents for the Web

Intelligent Information Retrieval 3 IR on the Web vs. Classsic IR  Input: publicly accessible Web  Goal: retrieve high quality pages that are relevant to user’s need  static (text, audio, images, etc.)  dynamically generated (mostly database access)  What’s different about the Web:  large volume  distributed data  Heterogeneity of the data  lack of stability  high duplication  high linkage  lack of quality standard

Intelligent Information Retrieval 4 Search Engine Early History  In 1990, Alan Emtage of McGill Univ. developed Archie (short for “archives”)  Assembled lists of files available on many FTP servers.  Allowed regex search of these file names.  In 1993, Veronica and Jughead were developed to search names of text files available through Gopher servers.  In 1993, early Web robots (spiders) were built to collect URL’s:  Wanderer  ALIWEB (Archie-Like Index of the WEB)  WWW Worm (indexed URL’s and titles for regex search)  In 1994, Stanford grad students David Filo and Jerry Yang started manually collecting popular web sites into a topical hierarchy called Yahoo.

Intelligent Information Retrieval 5 Search Engine Early History  In early 1994, Brian Pinkerton developed WebCrawler as a class project at U Wash.  Eventually became part of Excite and AOL  A few months later, Fuzzy Maudlin, a grad student at CMU developed Lycos  First to use a standard IR system  First to index a large set of pages  In late 1995, DEC developed Altavista  Used a large farm of Alpha machines to quickly process large numbers of queries  Supported Boolean operators, phrases in queries.  In 1998, Larry Page and Sergey Brin, Ph.D. students at Stanford, started Google  Main advance was use of link analysis to rank results partially based on authority.

Intelligent Information Retrieval 6 Web Search Query String IR System Ranked Documents 1. Page1 2. Page2 3. Page3. Document corpus Web Spider

Intelligent Information Retrieval 7 Spiders (Robots/Bots/Crawlers)  Start with a comprehensive set of root URL’s from which to start the search.  Follow all links on these pages recursively to find additional pages.  Index all novel found pages in an inverted index as they are encountered.  May allow users to directly submit pages to be indexed (and crawled from).

Intelligent Information Retrieval 8 Search Strategy Trade-Off’s  Breadth-first search strategy explores uniformly outward from the root page but requires memory of all nodes on the previous level (exponential in depth). Standard spidering method.  Depth-first search requires memory of only depth times branching-factor (linear in depth) but gets “lost” pursuing a single thread.  Both strategies implementable using a queue of links (URL’s).

Intelligent Information Retrieval 9 Avoiding Page Duplication  Must detect when revisiting a page that has already been spidered (web is a graph not a tree).  Must efficiently index visited pages to allow rapid recognition test.  Tree indexing (e.g. trie)  Hashtable  Index page using URL as a key.  Must canonicalize URL’s (e.g. delete ending “/”)  Not detect duplicated or mirrored pages.  Index page using textual content as a key.  Requires first downloading page.

Intelligent Information Retrieval 10 Spidering Algorithm Initialize queue (Q) with initial set of known URL’s. Until Q empty or page or time limit exhausted: Pop URL, L, from front of Q. If L is not an HTML page (.gif,.jpeg,.ps,.pdf,.ppt…) continue loop. If already visited L, continue loop. Download page, P, for L. If cannot download P (e.g. 404 error, robot excluded) continue loop. Index P (e.g. add to inverted index or store cached copy). Parse P to obtain list of new links N. Append N to the end of Q.

Intelligent Information Retrieval 11 Queueing Strategy  How new links added to the queue determines search strategy.  FIFO (append to end of Q)  gives breadth-first search.  LIFO (add to front of Q)  gives depth-first search.  Heuristically ordering the Q gives a “focused crawler” that directs its search towards “interesting” pages.  May be able to use standard AI search algorithms such as Best-first search, A*, etc.

Intelligent Information Retrieval 12 Restricting Spidering  Restrict spider to a particular site.  Remove links to other sites from Q.  Restrict spider to a particular directory.  Remove links not in the specified directory.  Obey page-owner restrictions  robot exclusion protocol

Intelligent Information Retrieval 13 Anchor Text Indexing  Extract anchor text (between and ) of each link:  Anchor text is usually descriptive of the document to which it points.  Add anchor text to the content of the destination page to provide additional relevant keyword indices.  Used by Google:  Evil Empire  IBM  Helps when descriptive text in destination page is embedded in image logos rather than in accessible text.  Many times anchor text is not useful:  “click here”  Increases content more for popular pages with many in- coming links, increasing recall of these pages.  May even give higher weights to tokens from anchor text.

Intelligent Information Retrieval 14 Multi-Threaded Spidering  Bottleneck is network delay in downloading individual pages.  Best to have multiple threads running in parallel each requesting a page from a different host.  Distribute URL’s to threads to guarantee equitable distribution of requests across different hosts to maximize through-put and avoid overloading any single server.  Early Google spider had multiple coordinated crawlers with about 300 threads each, together able to download over 100 pages per second.

Intelligent Information Retrieval 15 Directed/Focused Spidering  Sort queue to explore more “interesting” pages first.  Two styles of focus:  Topic-Directed  Link-Directed

Intelligent Information Retrieval 16 Topic-Directed Spidering  Assume desired topic description or sample pages of interest are given.  Sort queue of links by the similarity (e.g. cosine metric) of their source pages and/or anchor text to this topic description.  Preferentially explores pages related to a specific topic.

Intelligent Information Retrieval 17 Link-Directed Spidering  Monitor links and keep track of in-degree and out- degree of each page encountered.  Sort queue to prefer popular pages with many in- coming links (authorities).  Sort queue to prefer summary pages with many out- going links (hubs).

Intelligent Information Retrieval 18 Keeping Spidered Pages Up to Date  Web is very dynamic: many new pages, updated pages, deleted pages, etc.  Periodically check spidered pages for updates and deletions:  Just look at header info (e.g. META tags on last update) to determine if page has changed, only reload entire page if needed.  Track how often each page is updated and preferentially return to pages which are historically more dynamic.  Preferentially update pages that are accessed more often to optimize freshness of more popular pages.

Intelligent Information Retrieval 19 Quality and the WWW The Case for Connectivity Analysis  Basic Idea: mine hyperlink information on the Web  Assumptions:  links often connect related pages  a link between pages is a “recommendation”  Approaches  classic IR: co-citation analysis (a.k.a. “bibliometrics”)  connectivity-based ranking (e.g., GOOGLE)  HITS - hypertext induced topic search

Intelligent Information Retrieval 20 Co-Citation Analysis  Has been around since the 50’s (Small, Garfield, White & McCain)  Used to identify core sets of  authors, journals, articles for particular fields of study  Main Idea:  Find pairs of papers that cite third papers  Look for commonalities 

Intelligent Information Retrieval 21 Co-citation analysis (From Garfield 98) The Global Map of Science, based on co- citation clustering: Size of the circle represents number of papers published in the area; Distance between circles represents the level of co-citation between the fields; By zooming in, deeper levels in the hierarchy can be exposed. The Global Map of Science, based on co- citation clustering: Size of the circle represents number of papers published in the area; Distance between circles represents the level of co-citation between the fields; By zooming in, deeper levels in the hierarchy can be exposed.

Intelligent Information Retrieval 22 Co-citation analysis (From Garfield 98) Zooming in on biomedicine, specialties including cardiology, immunology, etc., can be viewed.

Intelligent Information Retrieval 23 Co-citation analysis (From Garfield 98)

Intelligent Information Retrieval 24 CiteSeer: A Web Agent for Citation Analysis (Bollacker, Lawrence, Giles ) The CiteSeer agent consists of three main components: (i) sub-agent to automatically locate and acquire publications, (ii) document parser and database creator, (iii) browser interface which supports search by keyword and browsing by citation links.

CiteSeer: A Web Agent for Citation Analysis

Intelligent Information Retrieval 27 Citations vs. Links  Web links are a bit different than citations:  Many links are navigational.  Many pages with high in-degree are portals not content providers.  Not all links are endorsements.  Company websites don’t point to their competitors.  Citations to relevant literature is enforced by peer-review.  Authorities  pages that are recognized as providing significant, trustworthy, and useful information on a topic.  In-degree (number of pointers to a page) is one simple measure of authority.  However in-degree treats all links as equal. Should links from pages that are themselves authoritative count more?  Hubs  index pages that provide lots of useful links to relevant content pages (topic authorities).

Intelligent Information Retrieval 28 Hypertext Induced Topic Search  Basic Idea: look for “authority” and “hub” web pages (Kleinberg 98)  authority: definitive content for a topic  hub: index links to good content  The two distinctions tend to blend  Procedure:  Issue a query on a term, e.g. “java”  Get back a set of documents  Look at the inlink and outlink patterns for the set of retrieved documents  Perform statistical analysis to see which patterns are the most dominant ones  Technique was initially used in IBM’s CLEVER system  can find some good starting points for some topics  doesn’t solve the whole search problem!  doesn’t make explicit use of content (so may result in “topic drift” from original query)

Intelligent Information Retrieval 29 Hypertext Induced Topic Search  Intuition behind the HITS algorithm  Authority comes from in-edges  Being a good hub comes from out-edges  Mutually re-enforcing relationship  Better authority comes from in-edges of good hubs  Being a better hub comes from out-edges of to good authorities HubsAuthorities A good authority is a page that is pointed to by many good hubs. A good hub is a page that points to many good authorities. Together they tend to form a bipartite graph A good authority is a page that is pointed to by many good hubs. A good hub is a page that points to many good authorities. Together they tend to form a bipartite graph

Intelligent Information Retrieval 30 HITS Algorithm  Computes hubs and authorities for a particular topic specified by a normal query.  1. First determine a set of relevant pages for the query called the base set (base subgraph) S.  For a specific query Q, let the set of documents returned by a standard search engine be called the root set R. Initialize S to R.  Add to S all pages pointed to by any page in R.  Add to S all pages that point to any page in R.  Analyze the link structure of the web subgraph defined by S to find authority and hub pages in this set. R S

Intelligent Information Retrieval 31 HITS – Some Considerations  Base Limitations  To limit computational expense:  Limit number of root pages to the top 200 pages retrieved for the query.  Limit number of “back-pointer” pages to a random set of at most 50 pages returned by a “reverse link” query.  To eliminate purely navigational links:  Eliminate links between two pages on the same host.  To eliminate “non-authority-conveying” links:  Allow only m (m  4  8) pages from a given host as pointers to any individual page.  Authorities and In-Degree  Even within the base set S for a given query, the nodes with highest in- degree are not necessarily authorities (may just be generally popular pages like Yahoo or Amazon).  True authority pages are pointed to by a number of hubs (i.e. pages that point to lots of authorities).

Intelligent Information Retrieval 32 HITS: Iterative Algorithm  Use an iterative algorithm to slowly converge on a mutually reinforcing set of hubs and authorities.  Maintain for each page p  S:  Authority score: a p (vector a)  Hub score: h p (vector h)  Initialize all a p = h p = 1  Maintain normalized scores:  Authorities are pointed to by lots of good hubs:  Hubs point to lots of good authorities:

Intelligent Information Retrieval 33 Illustrated Update Rules 2 3 a 4 = h 1 + h 2 + h h 4 = a 5 + a 6 + a 7

Intelligent Information Retrieval 34 HITS Iterative Algorithm Initialize for all p  S: a p = h p = 1 For i = 1 to k: For all p  S: (update auth. scores) For all p  S: (update hub scores) For all p  S: a p = a p /c c: For all p  S: h p = h p /c c: (normalize a) (normalize h)

Intelligent Information Retrieval 35 HITS Example D A B C E D A C B E A: [0.0, 0.0, 2.0, 2.0, 1.0] D A C B E H: [4.0, 5.0, 0.0, 0.0, 0.0] D A C B E Norm A: [0.0, 0.0, 0.67, , 0.33] D A C B E Norm H: [0.62, 0.78, 0.0, 0.0, 0.0] First Iteration Normalize: divide each vector by its norm (square root of the sum of the squares)

Intelligent Information Retrieval 36 HITS Algorithm  Let HUB[v] and AUTH[v] represent the hub and authority values associated with a vertex v  Repeat until HUB and AUTH vectors converge  Normalize HUB and AUTH  HUB[v] :=  AUTH[u i ] for all u i with Edge(v, u i )  AUTH[v] :=  HUB[w i ] for all u i with Edge(w i, v) AH v u1u1 u2u2 ukuk... w1w1 w2w2 wkwk

Intelligent Information Retrieval 37 Convergence  Algorithm converges to a fix-point if iterated indefinitely.  Define A to be the adjacency matrix for the subgraph defined by S.  A ij = 1 for i  S, j  S iff i  j  Authority vector, a, converges to the principal eigenvector of A T A  Hub vector, h, converges to the principal eigenvector of AA T  In practice, 20 iterations produces fairly stable results.

Intelligent Information Retrieval 38 HITS Results  Authorities for query: “Java”  java.sun.com  comp.lang.java FAQ  Authorities for query “search engine”  Yahoo.com  Excite.com  Lycos.com  Altavista.com  Authorities for query “Gates”  Microsoft.com  roadahead.com In most cases, the final authorities were not in the initial root set generated using Altavista. Authorities were brought in from linked and reverse-linked pages and then HITS computed their high authority score.

Intelligent Information Retrieval 39 HITS: Other Applications  Finding Similar Pages Using Link Structure  Given a page, P, let R (the root set) be t (e.g. 200) pages that point to P.  Grow a base set S from R.  Run HITS on S.  Return the best authorities in S as the best similar-pages for P.  Finds authorities in the “link neighbor-hood” of P. Similar Pages to “honda.com”: - toyota.com - ford.com - bmwusa.com - saturncars.com - nissanmotors.com - audi.com - volvocars.com

Intelligent Information Retrieval 40 HITS: Other Applications  HITS for Clustering  An ambiguous query can result in the principal eigenvector only covering one of the possible meanings.  Non-principal eigenvectors may contain hubs & authorities for other meanings.  Example: “jaguar”:  Atari video game (principal eigenvector)  NFL Football team (2 nd non-princ. eigenvector)  Automobile (3 rd non-princ. eigenvector)  An application of Principle Component Analysis (PCA)

Intelligent Information Retrieval 41 HITS: Problems and Solutions  Some edges are wrong (not “recommendations”)  multiple edges from the same author  automatically generated  spam Solution: weight edges to limit influence  Topic Drift  Query: jaguar AND cars  Result: pages about cars in general Solution: analyze content and assign topic scores to nodes

Intelligent Information Retrieval 42 Modified HITS Algorithm  Let HUB[v] and AUTH[v] represent the hub and authority values associated with a vertex v  Repeat until HUB and AUTH vectors converge  Normalize HUB and AUTH  HUB[v] :=  AUTH[u i ]. TopicScore[u i ]. Weight(v, u i ) for all u i with Edge(v, u i )  AUTH[v] :=  HUB[w i ]. TopicScore[w i ]. Weight(w i, v) for all u i with Edge(w i, v)  Topic score is determined based on similarity measure between the query and the documents

Intelligent Information Retrieval 43 PageRank  Alternative link-analysis method used by Google (Brin & Page, 1998).  Does not attempt to capture the distinction between hubs and authorities.  Ranks pages just by authority.  Applied to the entire Web rather than a local neighborhood of pages surrounding the results of a query.

Intelligent Information Retrieval 44 Initial PageRank Idea  Just measuring in-degree (citation count) doesn’t account for the authority of the source of a link.  Initial page rank equation for page p:  N q is the total number of out-links from page q.  A page, q, “gives” an equal fraction of its authority to all the pages it points to (e.g. p).  c is a normalizing constant set so that the rank of all pages always sums to 1.

Intelligent Information Retrieval 45 Initial PageRank Idea  Can view it as a process of PageRank “flowing” from pages to the pages they cite

Intelligent Information Retrieval 46 Initial PageRank Algorithm  Iterate rank-flowing process until convergence: Let S be the total set of pages. Initialize  p  S: R(p) = 1/|S| Until ranks do not change (much) (convergence) For each p  S: For each p  S: R(p) = cR´(p) (normalize)

Intelligent Information Retrieval 47 Sample Stable Fixpoint

Intelligent Information Retrieval 48 Linear Algebra Version  Treat R as a vector over web pages.  Let A be a 2-d matrix over pages where  A vu = 1/N u if u  v else A vu = 0  Then R = cAR  R converges to the principal eigenvector of A.

Intelligent Information Retrieval 49 Problem with Initial Idea  A group of pages that only point to themselves but are pointed to by other pages act as a “rank sink” and absorb all the rank in the system.  Solutions: Rank Score  Introduce a “rank source” E that continually replenishes the rank of each page, p, by a fixed amount E(p).

Intelligent Information Retrieval 50 PageRank Algorithm Let S be the total set of pages. Let  p  S: E(p) =  /|S| (for some 0<  <1, e.g ) Initialize  p  S: R(p) = 1/|S| Until ranks do not change (much) ( convergence ) For each p  S: For each p  S: R(p) = cR´(p) (normalize)

Intelligent Information Retrieval PageRank Example A B C  = 0.3 A C B Initial R: [0.33, 0.33, 0.33] R’(C): R(A)/2 + R(B)/ /3 R’(B): R(A)/ /3 R’(A): 0.3/3 A C B R’: [0.1, 0.595, 0.27] A C B R: [0.104, 0.617, 0.28] Normalization factor: 1/[R’(A)+R’(B)+R’(C)] = 1/0.965 First Iteration Only: before normalization: after normalization:

Intelligent Information Retrieval 52 Random Surfer Model  PageRank can be seen as modeling a “random surfer” that starts on a random page and then at each point:  With probability E(p) randomly jumps to page p.  Otherwise, randomly follows a link on the current page.  R(p) models the probability that this random surfer will be on page p at any given time.  “E jumps” are needed to prevent the random surfer from getting “trapped” in web sinks with no outgoing links.

Intelligent Information Retrieval 53 Speed of Convergence  Early experiments on Google used 322 million links.  PageRank algorithm converged (within small tolerance) in about 52 iterations.  Number of iterations required for convergence is empirically O(log n) (where n is the number of links).  Therefore calculation is quite efficient.

Intelligent Information Retrieval 54 Google Ranking  Complete Google ranking includes (based on university publications prior to commercialization).  Vector-space similarity component.  Keyword proximity component.  HTML-tag weight component (e.g. title preference).  PageRank component.  Details of current commercial ranking functions are trade secrets.

Intelligent Information Retrieval 55 Personalized PageRank  PageRank can be biased (personalized) by changing E to a non-uniform distribution.  Restrict “random jumps” to a set of specified relevant pages.  For example, let E(p) = 0 except for one’s own home page, for which E(p) =   This results in a bias towards pages that are closer in the web graph to your own homepage.  Similar personalization can be achieved by setting E(p) for only pages p that are part of the user’s profile.

Intelligent Information Retrieval 56 PageRank-Biased Spidering  Use PageRank to direct (focus) a spider on “important” pages.  Compute page-rank using the current set of crawled pages.  Order the spider’s search queue based on current estimated PageRank.

Intelligent Information Retrieval 57 Link Analysis Conclusions  Link analysis uses information about the structure of the web graph to aid search.  It is one of the major innovations in web search.  It is the primary reason for Google’s success.

Intelligent Information Retrieval 58 Behavior-Based Ranking  Emergence of large-scale search engines allow for mining aggregate behavior analysis to improving ranking.  Basic Idea:  For each query Q, keep track of which docs in the results are clicked on  On subsequent requests for Q, re-order docs in results based on click-throughs.  Relevance assessment based on  Behavior/usage  vs. content

Intelligent Information Retrieval 59 Query-doc popularity matrix B Queries Docs q j B qj = number of times doc j clicked-through on query q When query q issued again, order docs by B qj values.

Intelligent Information Retrieval 60 Vector space implementation  Maintain a term-doc popularity matrix C  as opposed to query-doc popularity  initialized to all zeros  Each column represents a doc j  If doc j clicked on for query q, update C j  C j +  q (here q is viewed as a vector).  On a query q’, compute its cosine proximity to C j for all j.  Combine this with the regular text score.

Intelligent Information Retrieval 61 Issues  Normalization of C j after updating  Assumption of query compositionality  “white house” document popularity derived from “white” and “house”  Updating - live or batch?  Basic assumption:  Relevance can be directly measured by number of click throughs  Valid?